Compare the models people actually argue about.
Keep the recommendation honest.
Use case and source filter stay visible up top. Edits, saves, follows, and links stay available without running the page.
Keep the same models, change the use case, and recompute the recommendation without resetting the page.
Keep the compare set tight. Add one model at a time, then prune aggressively.
Save, follow, or copy this exact setup only after the comparison looks right.
DeepSeek V4 Flash 0420 (Max) leads this compare set for coding copilot.
The main warning is thinner evidence on Coding.
- Shared tests
- 13
- Close calls
- 0
- Sources
- All public
Read this page as a workspace for choosing between top models, not as a universal crown. The current frame is Coding copilot under best for this use case.
- Why the leader is ahead
- Chat / text · Search / tool use
- Where the evidence is thin
- 0 close-call benchmarks and 119 missing missing leader rows keep this from reading as settled.
- What to do next
- Inspect decisive benchmarks first, then open the disagreement page or head-to-head pages if the top line still feels too narrow.
| Intelligence Index AA · index Text · Chat / text | 2477.3% exact aliasverified runtime Row details
| n/a |
| Input price AA · /1M input tokens Text · Chat / text | $0.1 /1M input tokens80.7% exact aliasverified runtime Row details
| n/a |
| Output price AA · /1M output tokens Text · Chat / text | $0.3 /1M output tokens84% exact aliasverified runtime Row details
| n/a |
| GPQA AA · % Text · Reasoning / math / science | 89.4%90.1% exact aliasverified runtime Row details
| n/a |
| Humanity's Last Exam AA · % Text · Reasoning / math / science | 34.8%83.8% exact aliasverified runtime Row details
| n/a |
| CritPt AA · % Text · Reasoning / math / science | 7.1%79.5% exact aliasverified runtime Row details
| n/a |
| SciCode AA · % Code · Coding | 45.3%44% exact aliasverified runtime Row details
| n/a |
| Terminal-Bench Hard AA · % Code · Coding | 35.6%88.9% exact aliasverified runtime Row details
| n/a |
| Tau2-Bench Telecom AA · % Text · Search / tool use | 95%95% exact aliasverified runtime Row details
| n/a |
| AA-Omniscience accuracy AA · % Text · Chat / text | 36.8%78.5% exact aliasverified runtime Row details
| n/a |
| AA-Omniscience non-hallucination AA · % Text · Chat / text | 3.9%4% exact aliasverified runtime Row details
| n/a |
| IFBench AA · % Text · Chat / text | 79.2%98.9% exact aliasverified runtime Row details
| n/a |
What is doing the visible work
- No decisive benchmark
- The shared tests are too split or too sparse for a clean separator.
What changes the winner
- DeepSeek V4 Flash 0420 (Max)
- 119 visible benchmark gaps still leave room for the result to move.
Where evidence is missing
How this weighting reads the field
The current decision mode is grounded in the Coding copilot preset. This keeps the compare page connected to a visible use case instead of an unspoken “overall winner” claim.